MNRAS 000, 1–16 (2017) Preprint 24 October 2018 Compiled using MNRAS LATEX style file v3.

Deep multi-survey classification of variable stars

C. Aguirre,1? K. Pichara,1,2,3 † I. Becker1 ‡

1 Computer Science Department, Pontificia Universidad Católica de Chile.
2 Millennium Institute of Astrophysics, Chile
3 Institute of Applied Computational Science (IACS), Harvard. Cambridge, MA, USA.

Accepted XXX. Received YYY; in original form ZZZ

arXiv:1810.09440v1 [astro-ph.IM] 21 Oct 2018

ABSTRACT
During the last decade, a considerable amount of effort has been made to classify
variable stars using different machine learning techniques. Typically, light curves are
represented as vectors of statistical descriptors or features that are used to train vari-
ous algorithms. These features demand big computational powers that can last from
hours to days, making impossible to create scalable and efficient ways of automatically
classifying variable stars. Also, light curves from different surveys cannot be integrated
and analyzed together when using features, because of observational differences. For
example, having variations in cadence and filters, feature distributions become biased
and require expensive data-calibration models. The vast amount of data that will
be generated soon make necessary to develop scalable machine learning architectures
without expensive integration techniques. Convolutional Neural Networks have shown
impressing results in raw image classification and representation within the machine
learning literature. In this work, we present a novel Deep Learning model for light
curve classification, mainly based on convolutional units. Our architecture receives
as input the differences between time and magnitude of light curves. It captures the
essential classification patterns regardless of cadence and filter. In addition, we intro-
duce a novel data augmentation schema for unevenly sampled time series. We tested
our method using three different surveys: OGLE-III; Corot; and VVV, which differ in
filters, cadence, and area of the sky. We show that besides the benefit of scalability, our
model obtains state of the art levels accuracy in light curve classification benchmarks.
Key words: light curves – variable stars – supervised classification – neural net –
deep learning

1 INTRODUCTION demand large computational resources and aim to represent

the most relevant characteristics of light curves. A lot of ef-
There has been a considerable amount of effort trying to au-
fort has been made to design these features, and the creation
tomate the classification of variable stars (Debosscher et al.
of new ones implies a lot of time and research. The future
2007; Sarro et al. 2009; Bloom & Richards 2011; Pichara
surveys in Astronomy demand new ways of extracting these
et al. 2012; Nun et al. 2015; Mackenzie et al. 2016; Pichara
features. One example of the huge amount of data is The
et al. 2016b; Benavente et al. 2017). Variable stars such as
Large Synoptic Survey Telescope (LSST) (Borne et al. 2007;
RR Lyrae, Mira, and Cepheids are important for distant
Abell et al. 2009) that will start operating on 2022. It is es-
ladder measurements as shown in Bloom & Richards 2011.
timated that the LSST will produce 15 to 30TB of data per
The ability to classify variable stars is closely related to the
night. New ways of treating this information have been pro-
way light curves are represented. One way is to create vec-
posed (Mackenzie et al. 2016; Valenzuela & Pichara 2017a;
tors of statistical descriptors, called features, to represent
Naul et al. 2017; Gieseke et al. 2017). Mackenzie et al. 2016’s
each light curve (Bloom & Richards 2011; Nun et al. 2015).
uses an unsupervised feature learning algorithm to classify
One of the most popular set of statistical features is pre-
variable stars. Valenzuela & Pichara 2017a perform unsuper-
sented in Nun et al. (2015), also known as FATS features
vised classification by extracting local patterns among light
(stands for Feature Analysis for Time Series). These vectors
curves and create a ranking of similarity. To extract such
patterns they use a sliding window as done in Mackenzie
et al. 2016.
? E-mail: claguirre@uc.cl
† E-mail: kpb@ing.puc.cl Either way, none of these methods can be applied im-
‡ E-mail: iebecker@uc.cl mediately in new surveys. The limited overlap and depth

© 2017 The Authors

2 Carlos Aguirre et al.
Figure 1. Comparison of RR Lyrae ab and Cepheid stars in OGLE-III, VISTA and Corot Survey respectively. Difference between
magnitude and cadence is shown.
coverage among surveys makes difficult to share data. The
difference in filter and cadence makes it even harder without
any transformation to the data. Figure 1 shows an example
of the complexity that exists among stars and surveys. Light
curves have a difference in magnitude and time, and most
of the time they are not human-eye recognizable, even by
experts. Since all the magnitudes are calibrated by using
statistics, it does not work correctly because of underlying
differences between surveys. Figure 2 shows a comparison
of statistical features of RR Lyrae ab stars using three dif-
ferent catalogs. To the best of our knowledge, little efforts
have been made to create invariant training sets within the
datasets. Benavente et al. 2017 proposed an automatic sur-
vey invariant model of variable stars transforming FATS sta-
tistical vectors (Nun et al. 2015) from one survey to another.
As previously mentioned, these features have the problem
of being computationally expensive and the creation of new
ones implies a lot of time and research. Therefore, there is a
necessity of faster techniques able to use data from different
surveys.
Artificial neural networks (ANNs) have been known for
decades (Cybenko 1989; Hornik 1991), but the vast amount
of data needed to train them made them infeasible in the
past. The power of current telescopes and the amount of
data they generate have practically solved the problem. The
improves in technology and the big amount of data makes
ANNs feasible for the future challenges in astronomy.
Artificial neural networks or deep neural networks cre-
ate their own representation by combining and encoding the
input data using non-linear functions (LeCun et al. 2015).
Depending on the number of hidden layers, the capacity
of extracting features improve, together with the need for
more data (LeCun et al. 2015). Convolutional neural net-
works (CNNs) are a particular type of neural network that
have shown essential advantages in extracting features from
images (Krizhevsky et al. 2012). CNNs use filters and con-
volutions that respond to patterns of different spatial fre-
quency, allowing the network to learn how to capture the
most critical underlying patterns and beat most of the clas-
sification challenges (Krizhevsky et al. 2012). Time series,
as well as in images, have also proven to be a suitable field
for CNNs (Zheng et al. 2014; Jiang & Liang 2016).
In this paper, we propose a convolutional neural net-
work architecture that uses raw light curves from different
surveys. Our model can encode light curves and classify be-
tween classes and subclasses of variability. Our approach
MNRAS 000, 1–16 (2017)
 Paper 3
does not calculate any of the statistical features such as the
ones proposed in FATS, making our model scalable to vast
amounts of data. In addition, we present a novel data aug-
mentation schema, specific for light curves, used to balance
the proportion of training data among different classes of
variability. Data augmentation techniques are widely used
in the image community (Krizhevsky et al. 2012; Dieleman
et al. 2015; Gieseke et al. 2017).
We present an experimental analysis using three
datasets: OGLE-III (Udalski 2004), VISTA (Minniti et al.
2010) and Corot (Baglin et al. 2002; Bordé et al. 2003). The
used datasets differ in filters, cadence and observed sky-area.
Our approach obtains comparative results with a Random
Forest (RF) classifier, the most used model for light curve
classification (Richards et al. 2011; Dubath et al. 2011; Long
et al. 2012; Gieseke et al. 2017), that uses statistical features.
Finally, we produce a catalog of variable sources by cross-
matching VISTA and OGLE-III and made it available for
the community. 1 .
The remainder of this paper is organized as follows: Sec-
tion 2 gives an account of previous work on variable stars
classification and convolutional neural networks. In Section 3
we introduce relevant background theory. Section 4 explains
our architecture and Section 5 describes the datasets that
are used in our experiments. Section 6 explains the modi-
fications that are made to the data and Section 7 gives an
account of the parameters that are used in the architecture.
Section 8 shows our experimental analysis and Section 8.1
presents a study of the time complexity of our model. Fi-
nally, in Section 9, the main conclusions and future work
directions are given.

2 RELATED WORK
As mention before, there has been huge efforts to classify
variable stars (Richards et al. 2011; Nun et al. 2015, 2014;
Huijse et al. 2014; Mackenzie et al. 2016; Pichara et al.
2016a; Valenzuela & Pichara 2017b). The main approach
has been the extraction of features that represent the infor-
mation of light curves. Debosscher et al. 2007 was the first
one proposing 28 different features extracted from the pho-
tometric analysis. Sarro et al. 2009 continue the work by
introducing the color information using the OGLE survey
and carrying out an extensive error analysis. Richards et al.
2011 use 32 periodic features as well as kurtosis, skewness,
standard deviation, and stetson, among others, for variable
stars classification. Pichara et al. 2012 improve quasars de-
tection by using a boosted tree ensembles with continuous
auto regressive features (CAR). Pichara & Protopapas 2013
introduce a probabilistic graphical model to classify variable
stars by using catalogs with missing data.
Kim et al. 2014 use 22 features for classifying classes
and subclasses of variable stars using random forest. Nun
et al. 2015 published a library that facilitates the extraction
of features of light curves named FATS (Feature Analysis
for Time Series). More than 65 features are compiled and
Figure 2. Comparison of FATS feature in RR Lyrae ab using his-
togram plots of stars using OGLE-III, Corot and Vista Surveys.
Every feature is shown with its relative importance in classifica-
tion as mention in Nun et al. 2015.
put together in a Python library2 . Kim & Bailer-Jones 2016
publish a library for variable stars classification among seven
classes and subclasses. The library extracts sixteen features
that are considered survey-invariant and uses random forest
for doing the classification process.
A novel approach that differs from most of the previ-
ous papers is proposed by Mackenzie et al. 2016. He face
the light curve representation problem by designing and im-
plementing an unsupervised feature learning algorithm. His
work uses a sliding window that moves over the light curve
and get most of the underlying patterns that represent ev-
ery light curve. This window extracts features that are as
good as traditional statistics, solving the problem of high
computational power and removing the human from the

1 Datasets will be available in

http://gawa.academy.org/profile/<authorusername>/. A Data 2 Information about the features and manuals of how to use them
Warehouse for astronomy (Machin et al. 2018) are available as well.

MNRAS 000, 1–16 (2017)

4 Carlos Aguirre et al.
pre-processing loop. Mackenzie et al. 2016 work shows that
automatically learning the representation of light curves is
possible.
Since every survey has different cadences and filters, re-
search has been done to learn how to transfer information
among them. The ability to use labeled information with-
out an extensive work accelerates the process of bringing
new labeled datasets to the community. Long et al. 2012
propose using noise and periodicity to match distributions
of features between catalogs. Also, show that light curves
with the same source and different surveys would normally
have different values for their features. Benavente et al. 2017
represent the joint distribution of a subset of FATS features
from two surveys and creates a transformation between them
by using a probabilistic graphical model with approximate
inference. Even though he use FATS for his experiments,
they can be easily changed to Mackenzie et al. 2016 fea-
tures, making the process even faster. Pichara et al. 2016b
present a meta-classification approach for variable stars. He
show an interesting framework that combines classifiers pre-
viously trained in different sets of classes and features. His
approach avoids to re-train from scratch on new classifica-
tion problems. The algorithm learns the meta-classifier by
using features, but including their computational cost in the
optimization of the model.
All methods mentioned above invest lots of efforts find-
ing a way to represent light curves. There have been several
works in deep learning where the network itself is the one
in charge of learning the representation of data needed for
classification. Baglin et al. 2002 use a vanilla neural network
for classifying microlensing light curves from other types of
curves such as variable stars. Belokurov et al. 2003 continue
the work presenting two neural networks for microlensing
detection. Krizhevsky et al. 2012 show the importance of
CNNs for image-feature-extraction, and use them to classify,
achieving impressive results. Zeiler & Fergus 2014 study the
importance of using filters in each convolutional layer and
explain the feature extraction using the Imagenet dataset.
Dieleman et al. 2015 apply galaxy morphology classification
using deep convolutional neural networks. Cabrera-Vives
et al. 2017 use a rotation-invariant convolutional neural net-
work to classify transients stars in the HITS survey. Mahabal
et al. 2017 transform light curves into a two-dimensional ar-
ray and perform classification with a convolutional neural
network.
CNNs not only work on images. Many studies have been
done using one-dimensional time series. Zheng et al. 2014
use them to classify patient’s heartbeat using a multi-band
convolutional neural network on the electrocardiograph time
series. Jiang & Liang 2016 create a decision maker for a
cryptocurrency portfolio using a CNN on the daily price
information of each coin.
3 BACKGROUND THEORY
In this section, we introduce the basics on both artificial
neural networks and convolutional neural networks, to gain
the necessary insights on how our method works.
Figure 3. The forming basic unit the perceptron. It combines the
inputs with their respective weights and apply a non-linear func-
tion to produce the output. A bias b0 is added to each perceptron
letting it change the starting point.
3.1 Artificial Neural Networks
Artificial neural networks (ANN) are computational mod-
els based on the structure and functions of biological ax-
ons (Basheer & Hajmeer 2000). The ability for axons to
transmit, learn and forget information has been the inspi-
ration for neural networks. ANNs are capable of extracting
complex patterns from the input data using nonlinear func-
tions and learn from a vast amount of observed data. ANNs
have been used in different areas such as speech recognition
(Graves et al. 2013; Xiong et al. 2016, 2017), image recogni-
tion (Krizhevsky et al. 2012; Szegedy et al. 2015; Ren et al.
2015), and language translation (Jean et al. 2014; Sutskever
et al. 2014) among others.
The basic forming unit of a neural network is the percep-
tron (as axon in biology). As shown in Figure 3 a perceptron
takes inputs and combine them producing one output. For
each of the input, it has an associative weight wi that rep-
resent the importance of that input, and a bias b0 is added
to each perceptron. The perceptron combines those inputs
with their respective weights in a linear form and then uses
a non-linear activation function to produce the output:
N
!
Õ
Out put = f wi · xi + b0 (1)
i=1
Where f is the activation function. Two of the most
widely used activation functions are the tanh and the sigmoid
function because of their space complexity. Moreover, relu
functions are widely used in convolutions as they avoid sat-
uration and are less computationally expensive.
1
sigmoid(x) = (2)
1 + e−x
e x − e−x
tan(x) = (3)
e x + e−x
relu(x) = max(0, x) (4)
The most basic neural network is the vanilla architec-
ture consisting of three layers: (i) the input layer, (ii) the
hidden layer and (iii) the output layer. As shown in Figure
4 the input of a perceptron is the output of the previous
MNRAS 000, 1–16 (2017)
 Paper 5

Figure 4. A vanilla neural network with an input, hidden and

output layer. Total number of hyperparameters 17.

one, except for the input layer that does not have any input
and the output layer that does not have any output. A fully
connected layer is when every neuron in one layer connects
to every neuron in the other one. The vanilla architecture Figure 5. Step by step of a convolution process. A sliding window
consists of two fully connected layers. and a moving step are applied to the data and given as an input
to the next layer.
The number of perceptrons for each layer depends on
the architecture chosen and therefore the complexity of the
model. A neural network can have hundreds, thousands or
the inverse of the gradient and a learning rate as shown in
millions of them. The experience of the team, as well as
Werbos 1990.
experimenting different architectures, is critical for choosing
Training artificial neural networks with backpropaga-
the number of layers, perceptrons for each one and the num-
tion can be slow. Many methods have been proposed based
ber of filters to be used. The number of hyperparameters is
on stochastic gradient descent (SGD) (Ruder 2016). The
mainly given by the weights in the architecture. The input
massive astronomical datasets make training infeasible in
layer is where we submit our data and has as many neurons
practice, and mini-batches are used to speed up the process
as our input does. The hidden layer is the one in charge of
(LeCun et al. 1998). A training epoch corresponds to a pass
combining the inputs and creating a suitable representation.
over the entire dataset, and usually, many epochs are needed
Finally, the number of neurons in the output layer is as many
to achieve good results. The way weights are updated can
classes we want to classify.
change as well. One of the most widely used optimizers has
Many architectures have been proposed for artificial
been Adam optimizer as describe in Kingma & Ba 2014. It
neural networks. The vanilla architecture can be modified
relies on the first moment (mean) and second moment (vari-
in the number of hidden layers and the number of per-
ance) of the gradient to update the learning rates. Ruder
ceptrons per layer. ANNs with one hidden layer using sig-
2016 present an overview of the different gradient descent
moid functions are capable of approximating any continuous
optimizers and the advantages and disadvantages for each
functions on a subset of Rn (Cybenko 1989). However, the
one.
number of neurons needed to do this increases significantly,
which could be computationally infeasible. Adding more lay-
ers with fewer perceptrons can achieve same results without
3.2 Convolutional Neural Nets
affecting the performance of the net (Hornik 1991). More
than three hidden layers are considered deep neural net- Convolutional neural networks (CNN) are a type of deep
works (DNN). DNNs extract information or features com- neural network widely used in images (Krizhevsky et al.
bining outputs from perceptrons, but the number of weights 2012; LeCun et al. 2015). It consists of an input and out-
and data needed to train them significantly increases (Le- put layer as well as several hidden layers different from fully
Cun et al. 2015). connected ones.
To train artificial neural networks we find the weights A convolutional layer is a particular type of hidden layer
that minimize a loss function. For classification purpose, used in CNNs. Convolutional layers are in charge of extract-
one of the most use loss functions is the categorical cross- ing information using a sliding window.As shown in Figure 5
entropy for unbalanced datasets (De Boer et al. 2005). Ini- the window obtains local patterns from the input and com-
tially, weights are chosen at random and are updated be- bines them linearly with its weights (dotted line). Then ap-
tween epochs. We compare the desired output with the ac- ply a nonlinear function and pass it to the next layer. The
tual one and pursue to minimize the loss function using back- sliding window moves and extracts local information using
propagation with any form of Stochastic Gradient Descent different inputs but with the same weights. The idea is to
(SGD) (Ruder 2016). Then we update each weight using specialize this window to extract specific information from

MNRAS 000, 1–16 (2017)

6 Carlos Aguirre et al.
local data updating its weights. The size of the window, as
well as the moving step, are chosen before running the ar-
chitecture. Each window corresponds to a specific filter. The
number of windows is chosen beforehand. The number of fil-
ters can be seen as the number of features we would like to
extract. Convolutions are widely used because of their ca-
pacity of obtaining features with their translation invariant
characteristic using shared weights (Krizhevsky et al. 2012).
Zeiler & Fergus 2014 study the importance of using filters
inside each convolutional layer and show the activation pro-
cess of using different filters on the Imagenet dataset.
After applying convolutional layers, a fully connected
layer is used to mix the information extracted by the con-
volutional layers. Fully-connected hidden layers are added
to create more complex representations. Finally, the output
layer has as many nodes as classes we need.
4 METHOD DESCRIPTION
We propose an architecture that can classify variable stars
using different surveys. We now explain each layer of our
architecture, depicted in Figure 6.
Our architecture transforms each light curve to a matrix
representation using the difference between points. We use
two convolutional layers for extracting the local patterns and
turn them into a flat layer. Two fully connected layers are
used, and an output layer is plugged at the end to perform
the classification. In the following subsections, we describe
and give insights on each of the layers.
4.1 Pre-processing
In this phase, light curves are transformed into a matrix.
Having a balanced dataset is critical for our purpose of
multi-survey classification. Therefore, we use NM ax as the
maximum number of stars we can extract per class and sur-
vey. Section 6.3 explains in detail the selection of the light
curves for the database.
We transform each of these light curves in a matrix rep-
resentation of size 2 × N where 2 corresponds to the number
of channels (time and magnitude) and N to the number of
points used per light curve. Figure 7 shows an example of a
light curve in a matrix representation.
To compare light curves between catalogs a reshape to
the matrix must be made. Light curves differ in magnitude
and time and for comparing them the difference between
observations was used. A matrix of size M ×2× N was created
where M, 2 and N corresponds to the number of light curves,
channels, and numbers of observations used. Figure 7 shows
an example of the transformation of a light curve. Section
6.1 explains in detail this part of the process.
4.2 First Convolution
We apply a convolutional layer to each of the channels in
separate branches with shared weights. We use a shared
convolutional layer to preserve the objective of integrating
datasets with different cadences. Shared layers mean that
each of the filters is the same on every tower. The number
of filters is given by S1 . We chose 64 filters, to match the
number of features presented in Nun et al. 2015.
Our convolution does not use a max-pooling layer as
done in Jiang & Liang 2016. We avoid max-pooling because
of the low cadence of Vista Survey and the detriment of
losing information. The step function sw was set to 2 or 12
days (considering OGLE cadence of 6 days as average) and
the sliding window tw was set to 42 points or 250 days as
done in Mackenzie et al. 2016; Valenzuela & Pichara 2017a.
4.3 Second Convolution
After applying one convolution, we employ another one to
mix and create more complex features. Jiang & Liang 2016
showed that using two convolutions achieves better results.
As the first convolution, the number of filters is given
by S2 , and it was established to be half of the filters used in
the first convolution.
4.4 Flatten Layer
After extracting the local patterns, we transform the last
convolution into a flatten layer as in Jiang & Liang 2016;
Zheng et al. 2014. Our layer combines its patterns afterwards
with a hidden layer in a fully connected way.
4.5 Hidden Layer
We use a hidden layer to combine our extracted patterns,
and the number of cells is given by ncells . After several ex-
periments, we realize that 128 cells generate the best results.
We perform many experiments using sigmoid, relu and tanh
activating functions. We obtain the best results using tanh
activation, as most of the deep learning literature suggests
(LeCun et al. 1998).
4.6 Softmax Layer
In the output layer, there is one node per each of the possible
variability classes. We test two different amount of classes:
one for 4 classes of variable stars and the other for 9 sub-
classes. We use a so f tmax function to shrink the output to
the [0, 1] range. We can interpret the numbers from the out-
put nodes as the probability that the light curve belongs to
the class represented by that node.
Finally, we minimize the average across training using
categorical cross entropy. We use categorical cross entropy as
our loss function as we obtained best results and the datasets
use are unbalanced.
5 DATA
We apply our method to variable star classification using
three different surveys: ”The Optical Gravitational Lensing
Experiment” (OGLE) (Udalski 2004), ”The Vista Variable
in the Via Lactea” (VVV) (Minniti et al. 2010) and ”Convec-
tion, Rotation and planetary Transit” (CoRot) (Baglin et al.
2002; Bordé et al. 2003). We select these surveys because of
their difference in cadence and filters. In the following sub-
sections, we explain each of these surveys in detail.
MNRAS 000, 1–16 (2017)
 Paper 7

Figure 6. The Convolutional Network Architecture for multi-survey.

5.1 OGLE-III Table 1. Class distribution of OGLE-III labeled set.

The Optical Gravitational Lensing Experiment III (OGLE-
Class name Abbreviation Num. of Stars
III) corresponds to the third phase of the project (Udalski
2004). Its primary purpose was to detect microlensing events Classical Cepheids CEP 8031
and transiting planets in four fields: the galactic bulge, the RR Lyrae RRLyr 44262
large and small Magellanic clouds and the constellation of Long Period Variables LPV 343816
Carina. Eclipsing Binaries ECL 55863
For our experiment, we use 451, 972 labeled light curves.
The cadence is approximately six days and in the experi- Subclass Name Abbreviation Num. of Stars
ments is considered our survey with medium cadence. The
First-Overtone 10
band used by the survey is infrared and visible. We discard Classical Cepheid CEP10 2886
the visible band because of the low number of observations Fundamental-Mode F
per star compared to the infrared band. The class distribu- Classical Cepheid CEPF 4490
tion is shown in Table 1. RR Lyrae ab RRab 31418
RR Lyrae c RRc 10131
Mira Mira 8561
Semi-Regular Variables SRV 46602
5.2 The Vista Variable in the Vı́a Láctea Small Amplitude Red Giants OSARGs 288653
The Visible and Infrared Survey Telescope (Vista) started Contact Eclipsing Binary EC 51729
working in February 2010 (Minniti et al. 2010). Its mission Semi-Detached
Eclipsing Binary nonEC 4134
was to map the Milky Way bulge and a disk area of the
center of the Galaxy.
To obtain labeled light curves from Vista, we cross-

MNRAS 000, 1–16 (2017)

8 Carlos Aguirre et al.

Table 2. Class distribution of VVV labelled set.

Class name Abbreviation Num. of Stars

Classical Cepheids CEP 36

RR Lyrae RRLyr 15228
Long Period Variables LPV 228606
Eclipsing Binaries ECL 2604

Subclass Name Abbreviation Num. of Stars

First-Overtone 10
Classical Cepheid CEP10 5
Fundamental-Mode F
Classical Cepheid CEPF 23
RR Lyrae ab RRab 10567
RR Lyrae c RRc 4579
Mira Mira 8445
Semi-Regular Variables SRV 37366
Small Amplitude Red Giants OSARGs 182795
Contact Eclipsing Binary EC 1818
Semi-Detached
Eclipsing Binary nonEC 786

Table 3. Class distribution of CoRoT labelled set.

Class name Abbreviation Num. of Stars

Classical Cepheids CEP 125

RR Lyrae RRLyr 509
Long Period Variables LPV 109
Eclipsing Binaries ECL 568

Subclass Name Abbreviation Num. of Stars

RR Lyrae ab RRab 28
RR Lyrae c RRc 481

Because of its early stage, just a few instances have

Figure 7. Each light curve is transformed using a matrix with two
channels, time and magnitude. For every channel, the maximum
number of points is N . Each light curve is transformed using the
difference between points and a new matrix is created with two
channels.
match the Vista catalog with OGLE-III. We found 246, 474
stars in total. The cadence of the observations are approx-
imately every eighteen days and is considered our survey
with low cadence. The band used by the survey is mainly
k ps , and the class distribution of the labeled subset is shown
in Table 2.
5.3 CoRoT
The Convection, Rotation and planetary Transits (CoRoT)
is a telescope launched in December 2016 (Baglin et al. 2002;
Bordé et al. 2003). Its main purpose is to continuously ob-
serve the milky way for periods up to 6 months and search
for extrasolar planets using transit photometry. One of the
main advantages is the high cadence that can be more than
a 100 observations per object per day.
been labeled. For our experiments, we use 1, 311 labeled light
curves. The cadence of the observations are approximately
every sixty per day and in the experiments is considered our
survey with a high cadence. This catalog does not use any
specific filter but the observations per object are in red, blue
and green bands and for the experiments, we used the white
band combining this three. The class distribution is shown
in Table 3
6 DATA PRE-PROCESSING
In this section, we explain in detail how to pre-process the
data to produce our architecture inputs. Given that we are
integrating several surveys that contain light curves with
different bands and number of observations, we have to pre-
process the data to get a survey-invariant architecture.
6.1 Time & Magnitude Difference
The difference between instruments, cadences and bands be-
tween surveys create a bias in the variability patterns of light
curves (Long et al. 2012; Benavente et al. 2017). We use as
an input the difference between the time and magnitude

MNRAS 000, 1–16 (2017)


Figure 8. Example of a new light curve using a burning param-
eter of 2 and a step parameter of 1.
with the previous measurements. This removes the extinc-
tion, distance and survey specific biases. Moreover, it acts
as a normalization method. It enables the network to learn
patterns directly from the observations without the need to
pre-processing any of the data or extinction correction.
6.2 Light Curve padding
The difference in cadence among OGLE-III, Vista, and
Corot catalogs create a big variance in the number of ob-
servations per light curve. To overcome this problem, we
impose a minimum number of observations and use a zero
padding to complete the light curves that cannot reach that
minimum. This is inspired by the padding procedure done
in deep learning for image analysis. To define such limit, we
tried many different values and notice that classification re-
sults do not change significantly within a range of 500 and
1, 500 observations. We fixed the limit at 500 points per light
curve because that amount preserves the classification ac-
curacy and keeps most of the light curves suitable for our
analysis.
6.3 A Light curve data augmentation model
Hensman & Masko 2015 studied the impact of unbalanced
datasets in convolutional neural networks and proved that
for better classification performance the dataset should be
class-balanced. Since the datasets used in this paper are un-
balanced, data augmentation techniques have to be applied
(Krizhevsky et al. 2012).
To balance the dataset, we propose a novel data aug-
mentation technique based on light curve replicates. As men-
tion before in Section 4, the number of stars per class and
survey is given by NM ax . If the number of light curves per
class and survey is larger than this parameter, the replica-
tion process does not take place. Otherwise, the light curves
are replicated until they reach that limit. Each class is repli-
cated using two light curve parameters: burning and step.
The burning parameter indicates how many points we have
to discard in the light curve. The step parameter tells ev-
ery how many points we should take samples. The burning
MNRAS 000, 1–16 (2017)
Paper 9
parameter goes up every time the light curve has been repli-
cated making the starting point different for each new data
of a specific class and survey. Figure 8 shows an example
of the replication of a light curve. Is important to note for
surveys with high and medium cadences such as Corot and
OGLE-III, the loss is not critical as many observations are
available. In cases of low cadence catalogs, such as Vista,
the loss of observations is significantly reduced depending
on the step parameter. To keep a minimum observation loss,
the step parameter is set to a random number between 0
and 2. The maximum replication of a light curve is 5.
7 PARAMETER INITIALIZATION
As in most of deep learning solutions, our architecture needs
the set up of several initial parameters. In this section, we
explain the model design and how to set up its parameters.
7.1 Parameters
As previously noted, surveys have different optics and ob-
servation strategies which impact the depth at which they
can observe. This impacts the number of variable stars de-
tected and cataloged. In our case, OGLE has been operating
longer and observes large portions of the sky while VVV goes
deeper but in a smaller area, and Corot has great time res-
olution but is considerably shallower. The combined catalog
is dominated by OGLE stars and the subclasses are highly
unbalanced, being the LPV class the majority of them.
In order to train with a more balanced dataset, we use
a limit of 8, 000 stars per class and survey. We test different
values and set it to 8, 000 as most of the classes and sub-
classes of VISTA and OGLE survey possess that amount as
shown in section 5. Finally, after several experiments mea-
suring the training speed and efficiency, we set the batch size
to 256. Table 4 shows a summary of the parameters of our
architecture.
7.2 Layers
We use two convolutional layers as done in Jiang & Liang
2016. In the imaging literature, several works show that one
convolutional layer is not enough to learn a suitable repre-
sentation, and commonly they use two convolutions (Zheng
et al. 2014; Jiang & Liang 2016; Gieseke et al. 2017).
We try using only one convolution and performance was
critically reduced. Three convolutions are also utilized, pro-
ducing results as good as using two, but the time for training
the net and the number of parameters increase significantly.
A window size tw was used for the convolution process
and set to 42 observations or 250 days in average as done in
Mackenzie et al. 2016; Valenzuela & Pichara 2017a. Finally, a
stride value sw was used and set to 2 or 12 days in average as
done in Mackenzie et al. 2016; Valenzuela & Pichara 2017a.
7.3 Activation functions
In convolutional layers we used relu activation function as
they are capable of extracting the important information .
relu(x) = max(0, x)) (5)
10 Carlos Aguirre et al.

Table 4. Parameters used for the proposed architecture

Parameter name Abbreviation Value

Global Parameters

Stars per survey and class NM a x 8000

Number of Points
Per Light Curve N 500
Batch Size - 256

Architecture

Filters for first convolution S1 64

Filters for second convolution S2 32
Window size tw 42
Stride Value sw 2
Perceptrons in the hidden layer nc ell s 128
Dropout - 0.5
Figure 10. Accuracy of the training model using a 10-fold strat-
Data Augmentation Parameters ified cross validation with subclasses.

Burning burning [1,5]

Step step [0,2]
Figure 9. Accuracy of the training model using a 10-fold strati-
fied cross validation with classes.
For hidden layers, except for convolutional, we used tanh
functions because of better results, there widely used and
better gradients they provide (LeCun et al. 1998).
7.4 Dropout
Dropout is a regularization technique which randomly drops
units in the training phase (Srivastava et al. 2014). Srivas-
tava et al. 2014 prove the importance of dropout in neural
networks to prevent overfitting. We used a dropout of 0.5
as suggested in the mentioned work in the two fully con-
nected parts of our architecture. One between the flattening
layer and hidden layer, the other between the hidden layer
and the output layer. Dropout increases the generalization
of the network, therefore the performance.
8 RESULTS
Our experiments are mainly intended to evaluate the classifi-
cation accuracy of the datasets described in Section 5 as well
as the time taken for training the model. We test our model
using classes and subclasses and compare it with a Random
Forest classifier (RF) (Breiman 2001), still the most used
model for light curve classification (Richards et al. 2011;
Dubath et al. 2011; Long et al. 2012; Gieseke et al. 2017).
In classification accuracy, our model achieves better results
in ECL classes and subclasses and comparable results in
LPV and RR Lyrae. RF produces better results in OGLE-
III dataset.
Fifty-nine features were extracted for each of the sur-
veys using FATS library, computed over 500 observations
per light curve. For more information about the features ex-
tracted please refer to Nun et al. 2015. We use the extracted
features to train the RF.
For all of the experiments, we perform a 10-fold strati-
fied cross-validation. The stratification is done by class and
survey to keep the same proportions of classes per survey in
the training and testing sets. We perform data augmenta-
tion on each of the steps of the cross-validation within the
training set. Data augmentation technique is important for
improving results in both models. Finally, a 20% validation
set is used in our model to control the number of epochs
during the training process.
8.1 Computational Run Time
As mentioned before, the LSST will start working in 2022
generating approximately 15TB per night. The vast amount
of data arriving in the future will demand scalable algo-
rithms. We measure the execution time for each of the
classification algorithms, considering the feature extraction
(needed just for the RF) and the training iterations. In Table
5 we can see that because of the feature-extraction required
in RF, the RF model takes way more time in overall. RF is
faster than CNN in the training phase because CNN needs
several epochs, in this case, 500 epochs, where each one takes
between 6 and 11 seconds to run for classes and subclasses
respectively. The extraction of features is parallelized using
CPUs, and the model training is parallelized using GPUs.

MNRAS 000, 1–16 (2017)

 Paper 11

Table 5. Approximately time of extraction of features and train- 8.2 Results with general classes of variability.
ing the algorithms.
We test our model using four general classes: (i) Cepheids
(CEP), (ii) Long Period Variables (LPV), (iii) RR Lyrae
Method Extraction Training Total
(RRlyr) and (iv) Eclipsing Binaries (ECL). The distribu-
of Features Algorithm Run Time
tion of classes and subclasses per survey are shown in Ta-
RF 11.5 days 36 min 11.52 days bles 1, 2 and 3. Figure 11 and 12 show the results of using
Classes CNN 30 min 50 min 1.33 hrs our convolutional architecture and RF respectively. Table 6
Subclasses CNN 30 min 91.8 min 2.03 hrs summarize the accuracy per class of both approaches.
As it can be seen, RF achieves 96% of accuracy in
Table 6. Accuracy per class and subclass for each survey. OGLE-III dataset as it has more labeled data than the oth-
ers surveys. In VVV RF obtains 97% of accuracy in some of
Class CNN RF the classes that have more labeled data (LPV ), but not in
ECL-OGLE 0.98 ± 0.01 0.97 ± 0.01 stars with few labeled examples (ECL and CEP ). In Corot,
ECL-VVV 0.92 ± 0.02 0.89 ± 0.03 RF achieves 91% of accuracy only in ECL stars, mainly be-
ECL-Corot 0.00 ± 0.00 0.91 ± 0.04 cause of the high cadence of Corot that make it infeasible
LPV-OGLE 0.99 ± 0.00 0.97 ± 0.01 to extract features correctly, especially those related to pe-
LPV-VVV 0.94 ± 0.01 0.97 ± 0.01 riodicity. RF results show that FATS features of some light
LPV-Corot 0.92 ± 0.11 0.00 ± 0.00 curves (LPV and RRlyr ) can classify accurately between dif-
RRLyr-OGLE 0.94 ± 0.01 0.97 ± 0.00 ferent surveys. That is not a surprise mainly because period
RRLyr-VVV 0.94 ± 0.01 0.86 ± 0.02 features are less sensitive to changes in cadence.
RRLyr-Corot 0.00 ± 0.00 0.58 ± 0.07
Our proposed architecture achieves comparable classi-
CEP-OGLE 0.90 ± 0.03 0.93 ± 0.01
fication accuracy in OGLE-III but with much less training
CEP-VVV 0.00 ± 0.00 0.08 ± 0.17
CEP-Corot 0.90 ± 0.08 0.46 ± 0.12 time. As shown in Figure 9 our model produces approxi-
mately an accuracy of 97% in the validation set. Each of the
Subclass CNN RF colors represents one training of the 10-fold stratified cross-
EC-OGLE 0.93 ± 0.01 0.91 ± 0.01 validation. As shown in Table 6, OGLE-III dataset achieves
EC-VVV 0.69 ± 0.04 0.71 ± 0.03 95% of accuracy in average in all of its classes. VVV survey
nonEC-OGLE 0.98 ± 0.01 0.94 ± 0.01 achieves 93.7% of accuracy in most of its classes, except for
nonEC-VVV 0.51 ± 0.04 0.37 ± 0.05 CEP stars, which are less than 40 light curves. Comparing it
Mira-OGLE 0.98 ± 0.01 0.98 ± 0.00 to RF performance, it achieves better performance in VVV
Mira-VVV 0.94 ± 0.01 0.29 ± 0.02 as it has 92% of accuracy on each of the classes, except for
SRV-OGLE 0.92 ± 0.02 0.93 ± 0.01 CEP. In Corot, CNN and RF achieve comparable results.
SRV-VVV 0.00 ± 0.00 0.67 ± 0.02
Osarg-OGLE 0.90 ± 0.01 0.88 ± 0.01
Osarg-VVV 0.00 ± 0.00 0.72 ± 0.01
RRab-OGLE 0.72 ± 0.03 0.96 ± 0.01 8.3 Results with subclasses of variability
RRab-VVV 0.77 ± 0.02 0.83 ± 0.01 We test our model using nine types of subclasses:
RRab-Corot 0.11 ± 0.15 0.48 ± 0.29
(i) First-Overtone 10 Classical Cepheid (Cep10), (ii) Funda-
RRc-OGLE 0.86 ± 0.02 0.98 ± 0.00
mental-Mode F Classical Cepheid (CepF), (iii) RR Lyrae ab
RRc-VVV 0.75 ± 0.03 0.83 ± 0.02
RRc-Corot 0.01 ± 0.01 0.99 ± 0.01 (RRab), (iv) RR Lyrae c (RRc), (v) Mira, (vi) Semi-Regu-
CEP10-OGLE 0.84 ± 0.03 0.92 ± 0.02 lar Variables (SRV), (vii) Small Amplitude Red Giants (OS-
CEP10-VVV 0.00 ± 0.00 0.00 ± 0.00 ARGs), (viii) Eclipsing Binaries (EC) and (ix) Eclipsing
CEPF-OGLE 0.72 ± 0.02 0.90 ± 0.01 Binaries Semi-Detached (nonEC). The distribution of sub-
CEPF-VVV 0.00 ± 0.00 0.00 ± 0.00 classes per survey are shown in Tables 1, 2 and 3. Figure 13
and 14 show the results of using our neural network archi-
tecture and Random Forest respectively. Table 6 summarize
the accuracy per subclass of both approaches.
As shown in Figure 14, RF achieves better accuracy in
RR Lyrae and Cepheid stars, despite the small number of
Note that a significant improvement in the feature extraction light curves and mainly because of the data augmentation
process could be made if FATS library supported GPUs. technique. As shown in Table 6, OGLE-III dataset achieves
Our proposed architecture and RF are trained using a more than 90% of accuracy in most of the subclasses, as it
computer with 128 GB RAM, a GeForce GTX 1080 Ti GPU has more labeled data than the others surveys. However, in
and 6 CPU. Our algorithm is developed using Keras (Chollet VVV survey, a 80% of accuracy is obtained in RRab and
et al. 2015) framework that runs on top of Tensorflow (Abadi RRc stars. Finally, Corot’s catalog achieves 99% of accuracy
et al. 2015) library. We use the scikit-learn (Pedregosa et al. in RRc stars.
2011) implementation for RF with defaults settings except As shown in Figure 10, our model produces approxi-
for the minimum samples leaf that was set to a 100 for better mately an accuracy of 85% in the validation set. In OGLE-III
accuracy. dataset, nonEC and Mira stars achieve a 98% of accuracy,
We can see that our method is significantly faster as it and 92% in EC, SRV, and Osarg stars. In VVV survey, our
works with raw magnitudes and time and requires only a model achieves 76% of accuracy in RRab and RRc stars, and
couple of minutes of pre-processing. 94% in Mira stars. However, Osarg and SRV are confused

MNRAS 000, 1–16 (2017)

12 Carlos Aguirre et al.
Figure 11. Confusion matrix per class and survey for the convolutional neural network. Empty cells correspond to 0%.
in a 81% and 93% with Mira type respectively, which indi-
cates a clear overfitting of LPV stars. None of the models
can correctly classify VVV Cepheids, mainly because of the
low number of light curves (28 light curves in total). With
EC classes from VVV we achieve better results than RF.
Finally, the accuracy achieved in Corot is the lowest, mainly
because of the few amounts of light curves used (28 RRab
and 481 RRc stars).
9 CONCLUSIONS
In this work, we have presented a CNN architecture to clas-
sify variable stars, tested on light curves integrated from
various surveys. The proposed model can learn from the se-
quence of differences between magnitude and time, automat-
ically discovering patterns across light curves even with odd
cadences and bands. We show that multi-survey classifica-
tion is possible with just one architecture and we believe it
deserves further attention soon. The proposed model is com-
parable to RF in classification accuracy but much better in
scalability. Also, our approach can correctly classify most of
the classes and subclasses of variability.
Like most of deep learning approaches, our model is
capable of learning its light curve representation, allowing
astronomers to use the raw time series as inputs.
In order to have an extra comparison point, we attempt
to compare our method with the approach presented in Ma-
habal et al. (2017). We implemented their algorithm; we
run it with the same catalogs we use for our method, tak-
ing 500 observations per light curve. Mahabal’s method did
MNRAS 000, 1–16 (2017)

Figure 12. Confusion matrix per class and survey using Random Forest algorithm. Empty cells correspond to 0%.
not generate results, because it takes about 5 hrs to process
just one light curve. The extra computational cost mainly
comes because their algorithm generates the 2D embeddings
by comparing every pair of points in each light curve, some-
thing that makes it impractical for our experimental setup.
To obtain results with Mahabal’s approach, we decrease the
number of observations per light curve until their method
return results in a comparable amount of time. With about
100 observations, their method takes about 5 minutes to
run, still too slow for our setup (about 140000 light curves,
8000 light curves per class, etc.) By using 50 observations,
their method could generate results. As we expected, their
classification results are much worse, given that the small
number of observations make impossible even capture the
light curve periods in most of the cases. We are not telling
MNRAS 000, 1–16 (2017)
Paper 13
that their method is worse than ours, it is just intended for
another problem setup. We believe that it is not a fear com-
parison to add these results to our paper, because Mahabal’s
method would be using much less information compared to
ours.
As future work, oversampling techniques should be
studied, given that we observe that our model is sen-
sitive to unbalanced training sets. Also, more complex
architectures must be developed to improve classification
accuracy. New approaches able to deal with few light
curves in some classes are needed. For example, simulation
models based on astrophysics would be a significant con-
tribution especially for the most unrepresented subclasses
of variability. With simulation models, deep learning
architectures could be significantly improved, given that
14 Carlos Aguirre et al.

Figure 13. Confusion matrix per class and survey for the convolutional neural network. Empty cells correspond to 0.

usually their performance is directly related to the number ACKNOWLEDGEMENTS

of training cases, even with simulated instances. In the
same direction, it would be interesting to produce larger
training sets by integrating a higher number of surveys.
Our code implementation is done in Python, available
for download at https://github.com/<authorusername>
/DeepMultiSurveyClassificationOfVariableStars. We
also published the catalogs with the cross-matched training We acknowledge the support from CONICYT-Chile,
sets used in this work. through the FONDECYT Regular project number 1180054.
This paper utilizes public information of the OGLE, Vista
and Corot surveys. The computations were made using the
cluster from Pontificia Universidad Católica de Chile.

MNRAS 000, 1–16 (2017)


Figure 14. Confusion matrix per class and survey using Random Forest algorithm. Empty cells correspond to 0.
REFERENCES
Abadi M., et al., 2015, TensorFlow: Large-Scale Machine Learning
on Heterogeneous Systems, https://www.tensorflow.org/
Abell P. A., et al., 2009, arXiv preprint arXiv:0912.0201
Baglin A., et al., 2002, in Stellar structure and habitable planet
finding. pp 17–24
Basheer I., Hajmeer M., 2000, Journal of microbiological methods,
43, 3
Belokurov V., Evans N. W., Du Y. L., 2003, Monthly Notices of
the Royal Astronomical Society, 341, 1373
Benavente P., Protopapas P., Pichara K., 2017, The Astrophysical
Journal, 845, 18pp
MNRAS 000, 1–16 (2017)
Paper 15
Bloom J., Richards J., 2011, Advances in Machine Learning and
Data Mining for Astronomy
Bordé P., Rouan D., Léger A., 2003, Astronomy & Astrophysics,
405, 1137
Borne K. D., Strauss M., Tyson J., 2007, BULLETIN-
AMERICAN ASTRONOMICAL SOCIETY, 39, 137
Breiman L., 2001, Machine learning, 45, 5
Cabrera-Vives G., Reyes I., Förster F., Estévez P. A., Maureira
J.-C., 2017, The Astrophysical Journal, 836, 97
Chollet F., et al., 2015, Keras, https://github.com/fchollet/
keras
Cybenko G., 1989, Mathematics of Control, Signals, and Systems
16 Carlos Aguirre et al.
(MCSS), 2, 303 Salakhutdinov R., 2014, Journal of machine learning research,
De Boer P.-T., Kroese D. P., Mannor S., Rubinstein R. Y., 2005, 15, 1929
Annals of operations research, 134, 19 Sutskever I., Vinyals O., Le Q. V., 2014, in Advances in neural
Debosscher J., Sarro L., Aerts C., Cuypers J., Vandenbussche B., information processing systems. pp 3104–3112
Garrido R., Solano E., 2007, Astronomy & Astrophysics, 475, Szegedy C., et al., 2015, in Proceedings of the IEEE conference
1159 on computer vision and pattern recognition. pp 1–9
Dieleman S., Willett K. W., Dambre J., 2015, Monthly notices of Udalski A., 2004, arXiv preprint astro-ph/0401123
the royal astronomical society, 450, 1441 Valenzuela L., Pichara K., 2017a, Monthly Notices of the Royal
Dubath P., et al., 2011, Monthly Notices of the Royal Astronom- Astronomical Society
ical Society, 414, 2602 Valenzuela L., Pichara K., 2017b, Monthly Notices of the Royal
Gieseke F., et al., 2017, Monthly Notices of the Royal Astronom- Astronomical Society
ical Society, p. stx2161 Werbos P. J., 1990, Proceedings of the IEEE, 78, 1550
Graves A., Mohamed A.-r., Hinton G., 2013, in Acoustics, speech Xiong W., Droppo J., Huang X., Seide F., Seltzer M., Stolcke A.,
and signal processing (icassp), 2013 ieee international confer- Yu D., Zweig G., 2016, arXiv preprint arXiv:1610.05256
ence on. pp 6645–6649 Xiong W., Droppo J., Huang X., Seide F., Seltzer M., Stolcke
Hensman P., Masko D., 2015, Degree Project in Computer Sci- A., Yu D., Zweig G., 2017, in Acoustics, Speech and Signal
ence, KTH Royal Institute of Technology Processing (ICASSP), 2017 IEEE International Conference
Hornik K., 1991, Neural networks, 4, 251 on. pp 5255–5259
Huijse P., Estevez P. A., Protopapas P., Principe J. C., Zegers P., Zeiler M. D., Fergus R., 2014, in European conference on com-
2014, IEEE Computational Intelligence Magazine, 9, 27 puter vision. pp 818–833
Jean S., Cho K., Memisevic R., Bengio Y., 2014, arXiv preprint Zheng Y., Liu Q., Chen E., Ge Y., Zhao J. L., 2014, in Interna-
arXiv:1412.2007 tional Conference on Web-Age Information Management. pp
Jiang Z., Liang J., 2016, arXiv preprint arXiv:1612.01277 298–310
Kim D.-W., Bailer-Jones C. A., 2016, Astronomy & Astrophysics,
587, A18 This paper has been typeset from a TEX/LATEX file prepared by
Kim D.-W., Protopapas P., Bailer-Jones C. A., Byun Y.-I., Chang the author.
S.-W., Marquette J.-B., Shin M.-S., 2014, Astronomy & As-
trophysics, 566, A43
Kingma D., Ba J., 2014, arXiv preprint arXiv:1412.6980
Krizhevsky A., Sutskever I., Hinton G. E., 2012, in Advances in
neural information processing systems. pp 1097–1105
LeCun Y., Bottou L., Orr G. B., Müller K.-R., 1998, in , Neural
networks: Tricks of the trade. Springer, pp 9–50
LeCun Y., Bengio Y., Hinton G., 2015, Nature, 521, 436
Long J. P., El Karoui N., Rice J. A., Richards J. W., Bloom J. S.,
2012, Publications of the Astronomical Society of the Pacific,
124, 280
Machin J. I., Pichara K., Protopapas P., Neyem A., Pieringer C.,
2018, Monthly Notices of the Royal Astronomical Society
Mackenzie C., Pichara K., Protopapas P., 2016, The Astrophysi-
cal Journal, 820, 138
Mahabal A., Sheth K., Gieseke F., Pai A., Djorgovski S. G., Drake
A., Graham M., et al., 2017, arXiv preprint arXiv:1709.06257
Minniti D., et al., 2010, New Astronomy, 15, 433
Naul B., Pérez F., Bloom J. S., van der Walt S., 2017, Nature
Astronomy, p. 1
Nun I., Pichara K., Protopapas P., Kim D.-W., 2014, The Astro-
physical Journal, 793, 23
Nun I., Protopapas P., Sim B., Zhu M., Dave R., Castro N.,
Pichara K., 2015, arXiv preprint arXiv:1506.00010
Pedregosa F., et al., 2011, Journal of Machine Learning Research,
12, 2825
Pichara K., Protopapas P., 2013, The Astrophysical Journal, 777,
83
Pichara K., Protopapas P., Kim D.-W., Marquette J.-B., Tis-
serand P., 2012, Monthly Notices of the Royal Astronomical
Society, 427, 1284
Pichara K., Protopapas P., Leon D., 2016a, The Astrophysical
Journal, 819
Pichara K., Protopapas P., León D., 2016b, The Astrophysical
Journal, 819, 18
Ren S., He K., Girshick R., Sun J., 2015, in Advances in neural
information processing systems. pp 91–99
Richards J. W., et al., 2011, The Astrophysical Journal, 733, 10
Ruder S., 2016, arXiv preprint arXiv:1609.04747
Sarro L., Debosscher J., López M., Aerts C., 2009, Astronomy &
Astrophysics, 494, 739
Srivastava N., Hinton G. E., Krizhevsky A., Sutskever I.,

MNRAS 000, 1–16 (2017)